FullAUC Optimization for Open-Set Recognition in Remote Sensing Images

To classify remote sensing images in the wild, open set recognition (OSR) is an important solution. It aims at classifying known classes of images and identifying the unknown classes that are not seen in the training stage. It needs to reduce the empirical and open space risk. The most prevalent prototype-based methods, which use very few prototypes to represent a class of images, may not fully capture the diversity of the images and hence increase open space risk. To address this issue, in this paper, we propose an OSR method, named full area under the receiver-operating-characteristic curve (FullAUC), and applied it to remote sensing image classification. FullAUC minimizes both of the risks in terms of AUC. The empirical risk is minimized by a multi-class AUC optimization algorithm on known classes. The open space risk is minimized by a binary-class AUC optimization, which discriminates all known classes against unknown classes. The advantage of FullAUC lies in that, by leveraging AUC-based ranking loss on data-to-data relations, FullAUC better captures data diversity than prototype-based methods. Another novel contribution is that we applied the background classes as a novel regularization strategy to remote sensing image classification, where background classes have been proven to be helpful in improving the generalization ability in the general OSR studies. Finally, to further enhance the detection of unknown classes, we employ a joint confidence-based decision rule with a sigmoid output layer, which mitigates the closed-set limitations of softmax. Experiments on several remote sensing benchmark datasets demonstrate the effectiveness of the proposed method. The source code is available at https://github.com/zjyellow/FullAUC.